machine learning algorithms for modelling consumption in ... · national) –let’s evaluate with...
TRANSCRIPT
5th International Conference on Smart Energy Systems
Copenhagen, 10-11 September 2019
#SESAAU2019
Machine learning algorithms for modelling consumption in district
heating systems
Danica MaljkovićEnergy Institute Hrvoje Požar
Zagreb, Croatia
Powered by 1
5th International Conference on Smart Energy Systems
Copenhagen, 10-11 September 2019
#SESAAU2019
Motivation & Background
• DH systems introduced mandatory individual metering in accordance with EU EED Art 14 resulting in significant dissatisfaction with a large share of final consumers
• Stipulations of the Law on Heat Market were halted from 2016 (!?) – still 30% of final consumers have not introduced individual metering (!?!?)
• Doubt that individual metering results in savings (apartment, building, network, national) – let’s evaluate with AI
• Twofold application – prediction and evaluation of energy efficiency measures
Powered by 2
5th International Conference on Smart Energy Systems
Copenhagen, 10-11 September 2019
#SESAAU2019
Heat Consumption (in DH)• Depends on a number of parameters of which some measurable and predictable, while other
are hard to predict• Measurable/predictable:
– building envelope characteristics– heating degree days (climatology)– number of occupants– schedule of space usage– energy source– building heat installation losses (including heating substation)– position of the apartment in the building– formula for calculating consumption in case of HCA– existence of individual metering
• Hard to predict: – heat gains/losses form adjacent apartments– heat comfort level in the apartment– mode od space usage (opening windows)– income level of the owner, readiness to pay
Powered by 3
5th International Conference on Smart Energy Systems
Copenhagen, 10-11 September 2019
#SESAAU2019
Data ?• Most accessible and reliable = billing data
• Time series: from 2010 to Dec 2018 (individual metering introduced in 70% of buildings in 2015)
• Additionally – questionnaire with behavioral data
• Data frame matrix of ~5,000,000 observations in 50 predictors
• Statistical learning with machine learning algorithms:Linear Regression, Logistic Regression, Decision Trees, Random Forest, Support Vector Machines (for now)
Powered by 4
5th International Conference on Smart Energy Systems
Copenhagen, 10-11 September 2019
#SESAAU2019
Did we save? Yes…
Powered by 5
5th International Conference on Smart Energy Systems
Copenhagen, 10-11 September 2019
#SESAAU2019
…but… (on the apartment level):
Powered by 6
5th International Conference on Smart Energy Systems
Copenhagen, 10-11 September 2019
#SESAAU2019
Building level?
Powered by 7
Reduced, as well,but less steep…
5th International Conference on Smart Energy Systems
Copenhagen, 10-11 September 2019
#SESAAU2019
Here come the outliers
Powered by 8
5th International Conference on Smart Energy Systems
Copenhagen, 10-11 September 2019
#SESAAU2019
Data distribution seems more “normal”
Powered by 9
5th International Conference on Smart Energy Systems
Copenhagen, 10-11 September 2019
#SESAAU2019
Network / Town Level
Powered by 10
5th International Conference on Smart Energy Systems
Copenhagen, 10-11 September 2019
#SESAAU2019
Learning and predicting
1. Take a portion of the data set (60-80%) and learn on it.
2. Test what you have learned on the remaining portion (20-40%). If accuracy level is satisfactory:
3. Predict on the “new” data and use the model.
Powered by 11
5th International Conference on Smart Energy Systems
Copenhagen, 10-11 September 2019
#SESAAU2019
Foresting the District Heating
Powered by 12
• Random Forest – best fit for (green) DH
• Pros: – great prediction
performance, robust, good variable importance estimate
• Cons: – less interpretable,
computationaly intensive
Predictor 1
?
Predictor 2
Predictor N
Result
Yes No
>40 <40
…….
…….
5th International Conference on Smart Energy Systems
Copenhagen, 10-11 September 2019
#SESAAU2019
Fitting of new data - prediction
Powered by 13
5th International Conference on Smart Energy Systems
Copenhagen, 10-11 September 2019
#SESAAU2019
Fitting accuracy is quite high
Type of random forest: regression
Number of trees: 500
No. of variables tried at each split: 8
Mean of squared residuals: 3075.775
(~55 kWh error)
% Var explained: 99.44
Powered by 14
5th International Conference on Smart Energy Systems
Copenhagen, 10-11 September 2019
#SESAAU2019
Importance plot
Powered by 15
5th International Conference on Smart Energy Systems
Copenhagen, 10-11 September 2019
#SESAAU2019
Predicting
Powered by 16
5th International Conference on Smart Energy Systems
Copenhagen, 10-11 September 2019
#SESAAU2019
Conclusions & Further Steps• Building energy consumption modelling and the usual methods
used are traditional multiple regression models, simulation methods and methods of artificial neural networks
• In the last years machine learning has gain larger application in various fields – „chewing of data”
• Gives good results in cases where large amounts of data are to be processed with an aim to recognize a pattern and correlation of each of the relevant parameter as well as in the cases where the problem is too complex for a human intelligence to solve
• Testing the results on other data sets, networks, cities, climates• Important variables crucial for determing policy in energy efficiency• Deep learning
Powered by 17
5th International Conference on Smart Energy Systems
Copenhagen, 10-11 September 2019
#SESAAU2019
QUESTIONS & THOUGHTS?
Danica Maljković
Energy Institute Hrvoje Pozar
Zagreb, Croatia
Powered by 18